Claude Code + Codex + opencode + any agent harness

Agent Workflow Kit

A portable setup pattern for turning one-off coding chats into repeatable workflows: define the playbook, run from a task queue, and preserve only the evidence needed for the next agent.

3core building blocks
1claimed task first
4handoff coordinates
anyagent harness
01
1Define
START
Write The
Playbook
Start by making the workflow a file, not a remembered chat habit. The file tells the agent how to choose work, what to read, what not to touch, and when to stop.
Portable entry points
CC
Claude Code command
Create a project command such as .claude/commands/dotodo.md so someone can type one slash command instead of pasting a long prompt.
CX
Codex instruction layer
Put repo conventions in AGENTS.md, then keep workflow-specific contracts in smaller referenced files so every task starts with the same rules.
OC
opencode agent
Define a primary or subagent with focused permissions, such as read-only planning, build work, review, or deployment handoff.
same idea, different harness file
AGENTS.mdportable repo rules
commandsrepeatable prompts
Keep it thinroute to detail files instead of pasting everything
02
2Run
Use A
Task Queue
The agent should not roam the repo looking for something useful. It should claim one real task, hydrate only the context needed for that task, and verify before it hands off.
Execution gates
1
Claim one task Linear, GitHub Issues, a local SQLite queue, or a JSON file can work if it has status, priority, owner, and notes.
CLAIM
2
Read progressively Start with task text, repo rules, recent notes, named files, and nearby tests before broader project scans.
NARROW
3
Split risk from work Planning, implementation, review, and landing can be separate commands or agents with different permissions.
BOUND
4
Stop on ambiguity If the task is unsafe, stale, blocked, or too large, update the queue instead of guessing in chat.
BLOCK
Good run
One task, one branch or change set, one verification story.
Bad run
A broad chat that mixes planning, coding, deploys, and unclear follow-ups.
03
3Reuse
Leave
Coordinates
The next agent should not need the whole conversation. It needs a short handoff, artifact paths, task ids, PR links, verification status, and known caveats.
Reusable artifacts
A
Compact state snapshot
Repo status, open PRs, claimed task, recent memory, and deploy state become one artifact instead of repeated discovery.
B
Evidence appendix
Long logs, screenshots, diffs, and command output are saved by path with a decisive summary in the handoff.
C
Fresh-session prompt
Landing, deploy, or follow-up work starts from coordinates, not a transcript copied across every turn.
short memory beats long transcripts
What to preserve
Safety !
Destructive and production boundaries
Never save tokens by hiding deletion rules, secret handling, or deploy approvals.
Proof V
Tests and review evidence
Keep pass/fail status, exact commands, screenshots, and skipped checks visible.
Next N
The next safe action
Close with the task id, PR URL, default action, and where to re-open evidence.